The Private Grid: When the Data Center Builds Its Own Power Plant

The cleanest way to see what the AI build-out actually is — not the brochure, the thing itself — is to stand near an xAI data center in the Mississippi Delta and count the turbines. By TechCrunch's reporting this spring, there were nearly fifty of them: trailer-mounted gas turbines, humming in the open air, generating electricity for a computer. The company's Memphis-area sites ran on the same setup — a cluster of gas-burning turbines that let the facility bypass the grid entirely while it waited on a connection from the Tennessee Valley Authority, power delivered by a contractor called VoltaGrid. At one point roughly thirty-five turbines produced around 422 megawatts on-site, with more than 500 megawatts deployed. That is not a backup generator. That is a private power plant, built in a hurry, in a neighborhood.

Why build your own power plant instead of plugging in? Because plugging in takes years. Lawrence Berkeley National Lab tracks the interconnection queue — the line every new power source waits in to join the grid — and its 2025 findings are stark: a typical project reaching operation in 2024 spent about 55 months, roughly five years, in that queue, double the wait of a decade ago, with more than 2,000 gigawatts of capacity backed up behind it. An AI lab racing to train a model on a two-year competitive clock cannot wait five years for electrons. So it doesn't. It decouples.

The mechanism of the decoupling is the tell. To run those Memphis turbines without triggering the Clean Air Act's stationary-source permitting, they were reportedly classified as "non-road engines" — the regulatory category for mobile equipment — because they sat on trailers. The Southern Environmental Law Center's verdict was not subtle: it called the result "an illegal power plant." A federal EPA rule update in early 2026 made that path harder, and the company has since said it will pull some turbines as a permanent substation comes online. But the template is now industry-wide: Natural Gas Intel projects data centers will draw roughly 35 gigawatts of behind-the-meter power — their own, off-grid generation — by 2030. Meta and others are building the same escape hatch.

Did AI do this, or did we?

No model asked for a gas turbine. The private grid is the product of a very human arbitrage: the gap between how fast capital wants to move and how slowly public infrastructure is allowed to. Faced with a five-year queue, a regulated permitting process, and a competitive deadline, the rational corporate move is to route around all three — to reclassify the turbine, to contract the power privately, to treat the public grid as the fallback rather than the foundation. Each decision is legal-until-challenged and locally rational. Stacked together, they add up to a quiet secession: the most valuable industrial customers in a generation opting out of the shared system, keeping the uptime, and leaving the grid's reliability problem — and the tailpipe — to everyone still connected to it. The carbon-neutral white paper and the diesel-adjacent reality are written by the same company in the same fiscal year.

What we are not claiming: that on-site generation is always illegitimate (bridging a genuine queue with temporary power is a real engineering answer), or that these turbines are the largest source of AI's emissions (the grid power behind the meter usually is). The documented concern is narrower and sharper — that the permitting shortcut and the queue-jump let a private operator externalize the air-quality and reliability costs onto a specific, usually less-powerful, community, while the compute and the valuation stay private. The turbines went up in the Mississippi Delta, not in a boardroom's line of sight.

The interconnection queue is a public commons — everyone waits their turn so the grid stays stable for all of it. The private grid is what happens when the largest customer decides the commons is optional. Our Energy lane keeps the ledger on who is opting out, where, and who breathes the difference.